Deprecate OVTrainer

[nikita-savelyevv]

What does this PR do?

Deprecate OVTrainer and related logic as stated in #890.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[echarlaix]

Works for me, @AlexKoff88 feel free to merge

[AlexKoff88]

@echarlaix, we decided to deprecate and remove it due to low interest. However, we are working on QAT for LLMs which could be contributed as an example by inheriting and implementing the HF Trainer interface instead of bringing it to the core functionality of Optimum-Intel.

[AlexKoff88]

@ljaljushkin, please take a look as well.

[echarlaix]

@echarlaix, we decided to deprecate and remove it due to low interest. However, we are working on QAT for LLMs which could be contributed as an example by inheriting and implementing the HF Trainer interface instead of bringing it to the core functionality of Optimum-Intel.

Why deprecating the OVTrainer in this context, what would be the benefit of having in the examples and not in optimum-intel directly as currently the case ?

[AlexKoff88]

@echarlaix, we decided to deprecate and remove it due to low interest. However, we are working on QAT for LLMs which could be contributed as an example by inheriting and implementing the HF Trainer interface instead of bringing it to the core functionality of Optimum-Intel.

Why deprecating the OVTrainer in this context, what would be the benefit of having in the examples and not in optimum-intel directly as currently the case ?

The main reason is to get flexibility in this flow:

• In the example, we will show how to implement QAT Trainer for LLMs with NNCF so that it is PyTorch model in and out
• User can modify the Trainer from the example according to the model training pipeline it has. NNCF part remains the same.
• User decides how to use optimized model: go to OV IR, use torch.compile, executorch, export to ONNX for ORT OV EP, etc.